Spaces:
Running
Running
File size: 13,809 Bytes
dc71cad | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 | """
localisation/deberta_ranker.py
βββββββββββββββββββββββββββββββ
Stage 2 β DeBERTa-v3-small cross-encoder ranker.
Given a set of candidate files from Stage 1 (RRF fusion), this
re-ranks them using a fine-tuned DeBERTa-v3-small cross-encoder that
classifies (issue_text, file_summary) β relevant/not_relevant.
Cross-encoders are much more precise than bi-encoders because they see
both the query AND the document together β allowing full attention
across both. The trade-off is they can't be pre-indexed (must run at
query time), so we only apply them to the top-20 candidates from Stage 1.
Training data (for fine-tuning):
- Positive: (issue_text, gold_file_summary) β label=1
- Negative: (issue_text, random_file_summary) β label=0
- Hard negatives: BM25 top-20 files that are NOT the gold file β label=0
- Dataset built from SWE-bench Lite instances
This module has two modes:
1. inference_only: loads a pre-trained checkpoint and scores candidates
2. training: fine-tunes DeBERTa-v3-small on the SWE-bench training set
For Phase 3 we implement the inference path + training scaffold.
Fine-tuning happens in Phase 7 (after trajectory data is collected).
"""
from __future__ import annotations
import logging
import os
from dataclasses import dataclass
from pathlib import Path
from typing import Optional
logger = logging.getLogger(__name__)
# Default model β can be swapped for a fine-tuned checkpoint
DEFAULT_MODEL = "microsoft/deberta-v3-small"
# Max token lengths for cross-encoder input
MAX_QUERY_LEN = 256 # issue text tokens
MAX_DOC_LEN = 256 # file summary tokens
MAX_TOTAL_LEN = 512 # total cross-encoder input length
@dataclass
class RankedFile:
file_path: str
relevance_score: float # 0β1 probability of relevance
rank: int # final rank (1-indexed)
stage1_rank: int # rank before re-ranking
class DeBERTaRanker:
"""
Cross-encoder re-ranker using DeBERTa-v3-small.
Scores each (issue, file_summary) pair and re-orders Stage 1 candidates.
Falls back gracefully to Stage 1 ordering if model unavailable.
"""
def __init__(
self,
model_name_or_path: str = DEFAULT_MODEL,
device: str = "auto",
max_length: int = MAX_TOTAL_LEN,
):
self.model_name_or_path = model_name_or_path
self.max_length = max_length
self._model = None
self._tokenizer = None
self._device = self._resolve_device(device)
self._available = False
self._try_load()
def _resolve_device(self, device: str) -> str:
if device != "auto":
return device
try:
import torch
if torch.cuda.is_available():
return "cuda"
if hasattr(torch.backends, "mps") and torch.backends.mps.is_available():
return "mps"
except ImportError:
pass
return "cpu"
def _try_load(self) -> None:
"""Attempt to load the model β log a warning if unavailable."""
try:
from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch
logger.info(
"Loading DeBERTa ranker: %s on %s", self.model_name_or_path, self._device
)
self._tokenizer = AutoTokenizer.from_pretrained(self.model_name_or_path)
self._model = AutoModelForSequenceClassification.from_pretrained(
self.model_name_or_path, num_labels=2
)
self._model.to(self._device)
self._model.eval()
self._available = True
logger.info("DeBERTa ranker loaded successfully")
except Exception as e:
logger.warning(
"DeBERTa ranker not available (%s) β will use Stage 1 ordering as-is", e
)
def rerank(
self,
issue_text: str,
candidates: list[tuple[str, str]], # list of (file_path, file_summary)
top_k: int = 10,
batch_size: int = 16,
) -> list[RankedFile]:
"""
Re-rank candidates by relevance to issue_text.
Args:
issue_text: the GitHub issue description
candidates: list of (file_path, file_summary) from Stage 1
top_k: number of results to return
batch_size: inference batch size
Returns:
List of RankedFile sorted by relevance_score descending
"""
if not candidates:
return []
if not self._available:
logger.debug("DeBERTa unavailable β returning Stage 1 ordering")
return [
RankedFile(
file_path=fp,
relevance_score=1.0 / (i + 1), # inverse rank as score
rank=i + 1,
stage1_rank=i + 1,
)
for i, (fp, _) in enumerate(candidates[:top_k])
]
# Score all candidates
scores = self._score_batch(issue_text, candidates, batch_size)
# Sort by score descending
ranked = sorted(
zip(candidates, scores),
key=lambda x: -x[1],
)
return [
RankedFile(
file_path=fp,
relevance_score=float(score),
rank=i + 1,
stage1_rank=next(
(j + 1 for j, (p, _) in enumerate(candidates) if p == fp), -1
),
)
for i, ((fp, _), score) in enumerate(ranked[:top_k])
]
def _score_batch(
self,
issue_text: str,
candidates: list[tuple[str, str]],
batch_size: int,
) -> list[float]:
"""Run cross-encoder inference on all candidates in batches."""
import torch
import torch.nn.functional as F
truncated_query = issue_text[:500] # characters (tokenizer handles tokens)
scores = []
for i in range(0, len(candidates), batch_size):
batch = candidates[i: i + batch_size]
texts_a = [truncated_query] * len(batch)
texts_b = [summary[:500] for _, summary in batch]
encoded = self._tokenizer(
texts_a,
texts_b,
max_length=self.max_length,
padding=True,
truncation=True,
return_tensors="pt",
)
encoded = {k: v.to(self._device) for k, v in encoded.items()}
with torch.no_grad():
logits = self._model(**encoded).logits
probs = F.softmax(logits, dim=-1)
# Class 1 = relevant
batch_scores = probs[:, 1].cpu().tolist()
scores.extend(batch_scores)
return scores
# ββ Training scaffold βββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
class DeBERTaTrainer:
"""
Fine-tunes DeBERTa-v3-small on (issue, file_summary) pairs.
Training data format (JSONL):
{"query": "<issue text>", "document": "<file summary>", "label": 0|1}
Called in Phase 7 after collecting trajectory data from SWE-bench runs.
"""
def __init__(
self,
base_model: str = DEFAULT_MODEL,
output_dir: Path = Path("models/deberta_ranker"),
num_epochs: int = 3,
learning_rate: float = 2e-5,
batch_size: int = 16,
):
self.base_model = base_model
self.output_dir = Path(output_dir)
self.num_epochs = num_epochs
self.learning_rate = learning_rate
self.batch_size = batch_size
def prepare_training_data(
self,
swe_instances, # list of SWEInstance
file_symbols_map, # {instance_id: list[FileSymbols]}
hard_negatives_k: int = 5, # BM25 top-k non-gold as hard negatives
) -> list[dict]:
"""
Build training pairs from SWE-bench instances.
Strategy:
Positive: (issue, gold_file_summary) β label=1
Hard-neg: BM25 top-5 files that are NOT in the gold patch β label=0
Random-neg: random repo file β label=0 (1:2 pos:neg ratio)
"""
from localisation.bm25_retriever import BM25Retriever
import random
training_pairs = []
for inst in swe_instances:
file_symbols = file_symbols_map.get(inst.instance_id, [])
if not file_symbols:
continue
# Extract gold file paths from the patch
gold_files = _extract_files_from_patch(inst.patch)
# Build BM25 index for this repo
retriever = BM25Retriever()
retriever.index(file_symbols)
bm25_hits = retriever.query(inst.problem_statement, top_k=hard_negatives_k + 5)
fs_map = {fs.file_path: fs for fs in file_symbols}
for gold_fp in gold_files:
if gold_fp not in fs_map:
continue
# Positive pair
training_pairs.append({
"query": inst.problem_statement[:500],
"document": fs_map[gold_fp].summary_text[:500],
"label": 1,
"instance_id": inst.instance_id,
})
# Hard negatives
for hit in bm25_hits[:hard_negatives_k]:
if hit.file_path not in gold_files and hit.file_path in fs_map:
training_pairs.append({
"query": inst.problem_statement[:500],
"document": fs_map[hit.file_path].summary_text[:500],
"label": 0,
"instance_id": inst.instance_id,
})
logger.info(
"Training data: %d pairs (%d positive, %d negative)",
len(training_pairs),
sum(1 for p in training_pairs if p["label"] == 1),
sum(1 for p in training_pairs if p["label"] == 0),
)
return training_pairs
def train(self, training_data: list[dict]) -> None:
"""Fine-tune DeBERTa on the prepared training data."""
try:
from transformers import (
AutoTokenizer, AutoModelForSequenceClassification,
TrainingArguments, Trainer
)
import torch
from torch.utils.data import Dataset
except ImportError as e:
raise ImportError("Install transformers + torch for fine-tuning") from e
class PairDataset(Dataset):
def __init__(self, data, tokenizer, max_length):
self.data = data
self.tokenizer = tokenizer
self.max_length = max_length
def __len__(self): return len(self.data)
def __getitem__(self, idx):
item = self.data[idx]
enc = self.tokenizer(
item["query"], item["document"],
max_length=self.max_length,
padding="max_length", truncation=True,
return_tensors="pt",
)
return {
"input_ids": enc["input_ids"].squeeze(),
"attention_mask": enc["attention_mask"].squeeze(),
"labels": torch.tensor(item["label"], dtype=torch.long),
}
tokenizer = AutoTokenizer.from_pretrained(self.base_model)
model = AutoModelForSequenceClassification.from_pretrained(
self.base_model, num_labels=2
)
dataset = PairDataset(training_data, tokenizer, MAX_TOTAL_LEN)
train_size = int(0.9 * len(dataset))
val_size = len(dataset) - train_size
train_ds, val_ds = torch.utils.data.random_split(dataset, [train_size, val_size])
args = TrainingArguments(
output_dir=str(self.output_dir),
num_train_epochs=self.num_epochs,
per_device_train_batch_size=self.batch_size,
per_device_eval_batch_size=self.batch_size,
learning_rate=self.learning_rate,
eval_strategy="epoch",
save_strategy="epoch",
load_best_model_at_end=True,
metric_for_best_model="eval_loss",
logging_steps=10,
warmup_ratio=0.1,
)
trainer = Trainer(
model=model, args=args,
train_dataset=train_ds, eval_dataset=val_ds,
)
trainer.train()
trainer.save_model(str(self.output_dir))
tokenizer.save_pretrained(str(self.output_dir))
logger.info("DeBERTa ranker saved to %s", self.output_dir)
# ββ Metric helpers ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def recall_at_k(
predictions: list[str],
gold_files: list[str],
k: int,
) -> float:
"""Compute recall@k: fraction of gold files in top-k predictions."""
if not gold_files:
return 0.0
top_k_set = set(predictions[:k])
hits = sum(1 for gf in gold_files if gf in top_k_set)
return hits / len(gold_files)
def _extract_files_from_patch(patch: str) -> list[str]:
"""Extract list of files modified in a unified diff."""
import re
# Match '--- a/path/to/file.py' or '+++ b/path/to/file.py'
pattern = re.compile(r"^(?:\+\+\+|---)\s+(?:a/|b/)(.+?)(?:\s|$)", re.MULTILINE)
files = list(set(pattern.findall(patch)))
return [f for f in files if f and f != "/dev/null"]
|